Learning What to Hear: Boosting Sound-Source Association for Robust Audiovisual Instance Segmentation

Audiovisual instance segmentation (AVIS) aims to accurately localize and track sounding objects throughout video sequences. Existing methods suffer from visual bias stemming from two fundamental issues: uniform additive fusion prevents queries from specializing to different sound sources, while visual-only training objectives limit queries from converging to arbitrary salient objects. We propose Audio-Centric Query Generation (ACQG) using cross-attention mechanism, enabling each query to selectively attend to distinct sound sources and carry sound-specific priors into visual decoding. Additionally, we introduce Sound-Aware Ordinal Counting (SAOC) loss that explicitly supervises sounding object numbers through ordinal regression with monotonic consistency constraints, preventing visual-only convergence during training. Experiments on AVISeg benchmark demonstrate consistent improvements: +1.64 mAP, +0.6 HOTA, and +2.06 FSLA, validating that query specialization and explicit counting supervision are crucial for accurate audiovisual instance segmentation. Our code and models are available at https://github.com/jinbae-s/ACVIS